LongVideoBench - release-reported accuracy
Multimodal · 2024-07-22
LongVideoBench evaluates referring reasoning over long, subtitled videos of up to an hour with multiple-choice questions that point to specific moments. This row holds release-reported accuracy; sources use the 1,337-question validation split or an unstated split and their own frame budgets, so no item count is declared and the protocol is recorded per row. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| Kimi K2.5 | 79.8 |
| Gemini 3 Pro Preview | 77.7 |
| GPT-5.2 | 76.5 |
| Opus 4.5 | 67.2 |
| Qwen3 VL 235B A22B Thinking | 65.6 |
| Kimi-VL-A3B-Instruct | 64.5 |
| Nemotron Nano 12B V2 VL BF16 | 63.6 |
| GPT-4o Mini | 58.2 |
| Qwen2.5-VL-7B-Instruct | 56.0 |
| Gemma 3 12B IT | 51.5 |
| Llama 3.2 11B Vision Instruct | 45.5 |